Re: [PATCH] drm/nv84-: write fence value on exit, and restore value on init.

[Maarten Lankhorst <maarten.lankhorst-Z7WLFzj8eWMS+FvcfC7Uqw@public.gmane.org>]

Op 04-09-13 05:21, Ben Skeggs schreef:
> On Tue, Sep 3, 2013 at 12:31 AM, Maarten Lankhorst
> <maarten.lankhorst-Z7WLFzj8eWMS+FvcfC7Uqw@public.gmane.org> wrote:
>> This increases the chance slightly that recovery from lockup can happen
>> succesfully.
> I'd *really* love to see proof of this.  When channels die, all
> outstanding fences are marked as signalled.  This should do absolutely
> nothing...
nv84+ heavily rely on fences though, and a race like this is possible:
- channel 0 uses a bo from channel 1, queues a wait somewhere in the command stream for it.
- channel 1 dies cleanly, but userspace creates a new channel in its place, fence counter is reset to 0.
- channel 0 reaches the NV84_SUBCHAN_SEMAPHORE_TRIGGER.ACQUIRE_GEQUAL op, waits on fence in channel 1 to signal forever.

Channel 0 could be the global drm channel used for buffer moves, which would result in a hang. This may seem unlikely, but I believe that parallel piglit runs could trigger it.

If not, simply creating an operation that takes a few seconds in channel 0 and then queuing a command that uses a bo from channel 1 while chan1 is still busy, then deleting/recreating chan1 could trigger it.

~Maarten